Memory Kernel Parameters.

[joe_91]

Hi All:

I am attaching the current kernel parameters on our
Nclass server. The parameters were set when we had
9Gig of memory and now we have 20 Gig. I need to
review and adjust the kernel for the 20 Gig memory.
We are running Xemulator from this box to publish
Applications from the PC. I need your suggestions with
if possible an explanation as to why on the relevant
parameters. Please help. I know to start with, i have
to reduce the dbc_max value. Please help me with your
valuable suggestions.

Thanks
Joe.

[Paula J Frazer-Campbell]

Joe

Two obvious ones are:-

dbc_max_pct 20
dbc_min_pct 5


Change to:-

dbc_max_pct 10
dbc_min_pct 3

Changing and Monitoring changes to dbc_max_pct
sar -d

The %rcache will directly show the results in a reduction in the dbc_max_pct setting ??? less that 100 = paging . Paging to disk in itself is not a problem provided that it is keep low as disk memory is the slowest medium.

[Craig Rants]

I would look here for recommendations. Paula's comment is right in line and covered in this doc. It is from the HP performance guru.

GL,
C

[Craig Rants]

oops, a link would be good

http://www.hp-partners.com/tcl_public/html/technical_support/tuning.html

[A. Clay Stephenson]

Under 11.11, I have found that most boxes with lots of memory (and your qualifies) do well with about 800-900MB of buffer cache. My preference is to used fixed buffer cache and forget that dynamic stuff. I would set bufpages to about 204800 (800MB). The other thing you might need to do is either add more swapspace (which would probably require an increase in swchunk) OR make sure that swapmem_on=1 to enable pseudoswap. I certainly hope that you don't swap at all with 20GB. You almost certainly should avoid the old 2-3x swap for memory rule. You would probably do fine with a very small amount of swap (512MB or so) with pseudoswap enabled.

[MANOJ SRIVASTAVA]

Hi Joe

The parameters that effect memeory are :

bufpages
dbc_max
dbc_min
desfree
lotsfree
maxdsiz
maxssiz
maxtsiz
minfree
msgmax.msgmnb
nbuf
nclist
netmemmax
nfile
ninode
nproc
page_text_to_local
strmsgsz
swapmem_on
unloackable_mem

Since you are upgrading in the existing working env so I would definaltey keep the size on dynamic buffer at least similar hence change dbc_max and dbc_min to give me like 7% of 20 Gb , then ofcourse monitor others before chaging them

Manoj Srivastava

[Paula J Frazer-Campbell]

Hi Joe

The DBC_MAX and DBC_MIN would be all that I would change.

Other may disagree !!!!


Paula

[joe_91]

Hi All:

Thanks for the quick response. We are running 11.0 and targetting 400 simultaneous user sessions from the xemulation. We found out during the last load test per user memory usage was 50 Meg. Assuming we set the buffer cache, what are the other memory related parameters we have to set?

Thanks
Joe.

[Sandip Ghosh]

Joe,

Except the dbc_max_pct I would look for a increase in shmmax. Since you are having enough memory , be generous.

Sandip

[joe_91]

Hi Paula:

Your reccomendation of 10 and 3 means that we are talking between 2 gig and 600 meg of buffer cache. when i change these values how do i know that the new changes are in effect(not by looking by kmtune but from other aspects) ALSO i thought the HP recommendation was about 300-400 megs for the buffer cache so what would be the effect if i set max to 3 and min to 2.

Thanks
Joe.

[Sandip Ghosh]

you should look at sar -b output. The %rchace should be 95 to 100% in case of optimum buffer setting. You can decrease further and look for the output.

Sandip

[A. Clay Stephenson]

Forget what I said about pseudoswap, you haven't bought enough memory. If your numbers about 50MB per instance with 400 instances that's 20GB there - with no overhead for the kernel or other processes. You 50GB may be an overestimate because of shared text/shared library usage but still it appears that you should purchase additional memory. If all 400 users are active, the system will be paging out. You will need to enable additional swapspace. Given your application, the maxdsiz, maxtsiz, and maxssiz limits should be fine. In fact, this is a case where you should consider reducing the values (or doing it the .profiles via ulimit) so that a few rogue processes would not start eating resources and bringing your 400 users to a crawl.

[Paula J Frazer-Campbell]

Joe

You are right about the circa 400 meg,

I erred on the safe side - you have doubled your memory (roughly) and I halved the settings.

If you want to use this as an opportunity to tune as well as make allowances for your extra memory then :-


dbc_max_pct = 2 - 400 meg
dbc_min_pct =1 - 200 meg


use sar ???b

and watch the %rcache colum for figures below 100% -

That is:- the system was not 100% successful in finding what it wanted in the rcache and so therefore user the slower menium of disk.


05:00:00 bread/s lread/s %rcache bwrit/s lwrit/s %wcache pread/s pwrit/s
05:01:00 243 15477 98 4 128 97 0 318
05:02:00 172 11227 98 6 170 97 0 279
05:03:00 89 21358 100 5 123 96 0 284
05:04:00 258 19605 99 7 121 94 0 332
05:05:00 267 9727 97 2 136 98 0 345
05:06:00 150 10813 99 3 172 98 0 286
05:07:00 326 7702 96 5 218 98 0 302
05:08:00 297 13572 98 3 178 99 0 314
05:09:00 391 5470 93 6 83 93 0 282
05:10:00 413 3306 87 4 57 94 0 326
05:11:00 374 3620 90 3 61 95 0 287
05:12:00 215 3986 95 4 82 95 0 205
05:13:00 401 3264 88 2 59 97 0 307


Paula